On Tuesday, April 8, Sam Altman posted that Codex had crossed 3 million weekly active users. He marked the milestone by resetting the product's usage limits and promised he would keep resetting them every additional million users until Codex hit ten. Developers on X treated it as a victory lap.
Eight days later, the victory lap turned into a full strategic pivot. The Codex desktop app OpenAI pushed out on April 16 does not behave like an upgrade. It behaves like a replacement.
The previous version was a coding agent that lived inside its own window. The new one reaches out of that window, takes control of your Mac, opens your other apps, clicks around in them, and runs multiple specialized agents in parallel while you keep working in a different window. That description sounded familiar to the developers reading TechCrunch's coverage on Thursday morning. It is roughly what Anthropic's Claude Code has been doing for weeks.
The Desktop App Is Now a Workflow Orchestrator
The centerpiece of the update is a feature OpenAI calls Computer Use, and in Codex it means what it sounds like. The agent can now open any application on the user's desktop, read what is on screen, move the cursor, type, click, and chain those actions into multi-step tasks. On Mac, it can do this in the background while the developer continues using the machine for something else. VentureBeat described it as the agent "using all other apps on your computer."
Alongside computer use, Codex gained four new capabilities that change the shape of the product.
The first is an in-app browser. Codex can now open web pages inside its own window, issue commands against them, and, per OpenAI's documentation, carry out tasks on specific web applications. The company pitches the browser for front-end and game development, where seeing the rendered output inside the agent's own context is the difference between a useful workflow and a broken one.
The second is Memory, shipped in preview. It lets Codex recall previous work sessions, remember personal preferences and prior corrections, and carry context forward without a developer rewriting custom instructions in every new session.
The third is image generation, usable for product concepts, slide visuals, mockups, and placeholder imagery inside a coding flow.
The fourth is the most quietly important: more than 90 new plugins. The list includes Atlassian Rovo, CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, and Render. OpenAI frames the plugin surface as a packaging unit that bundles skills, app integrations, and MCP servers into one thing the agent can reach for. Last month Codex had a handful of connectors. This week it has a supply chain.
Subagents Are the Feature That Actually Matters
Computer use grabs the headlines, but the structural shift inside Codex is subagents.
A subagent is a specialized Codex instance spawned by the main agent to handle one piece of a larger task. Codex can now run those subagents in parallel and collect their results into a single response. OpenAI's developer documentation, published on the Codex subagents page, describes use cases like exploring a large codebase, triaging a JIRA ticket, or implementing a multi-step feature plan while the developer is still writing a commit message somewhere else.
In practice, that turns a single-threaded coding conversation into something closer to a team of interns reporting back. One subagent might read every call site of a function about to be renamed. Another might run the test suite against a speculative fix. A third might draft the release notes. The main agent stitches the results together.
OpenAI says subagent workflows are enabled by default in current Codex releases. That default is the tell. Subagents are not an opt-in power-user feature. They are the new execution model.
The Competitive Context Is Anthropic
TechCrunch framed the update as "OpenAI takes aim at Anthropic," and that is the accurate reading. The features Codex gained on Thursday read almost line-by-line against what Anthropic has shipped in Claude Code over the prior six weeks.
| Capability | Codex (April 16) | Claude Code |
|---|---|---|
| Desktop computer use | Yes, macOS background | Yes, added earlier in 2026 |
| In-app browser | Yes | Yes |
| Multi-agent / subagents | Yes, default on | Yes, via Coordinator Mode |
| Memory | Preview | Yes |
| Plugin ecosystem | 90+ new plugins | MCP-based, broad |
| Underlying model | GPT-5.4 family | Claude Opus 4.7 |
That table is not a coincidence. Anthropic's Claude Code has been the product of record for agentic coding among senior engineers for months, and internal OpenAI metrics described to The Tech Portal show the gap in developer preference was widening rather than closing. Anthropic's Opus 4.7, which beats GPT-5.4 on SWE-bench Verified, sits underneath Claude Code and gives it a genuine quality advantage on long-horizon refactors.
Codex's answer is not a better model. It is a wider surface. If your agent can touch your whole Mac, dispatch five subagents in parallel, and plug into your ticket tracker, your CI, your code review bot, and your calendar, the model's raw benchmark score matters less than the tool's reach.
The Economics Are the Reason the Timing Is This Aggressive
Codex growth has been extraordinary. OpenAI told developers that usage measured in tokens grew more than 70% month over month through Q1 2026, and that users had quintupled in the preceding three months. The product crossed 3 million weekly users on April 8. Altman posted that he would reset limits again at 4 million, 5 million, and every milestone after that.
The economic pressure underneath those numbers is less romantic. Coding agents are the highest-margin, highest-retention product category in consumer AI right now. Cursor, the AI-first IDE, hit $2 billion in annualized revenue in early 2026 selling essentially one workflow: write code with an assistant sitting next to you. Claude Code is the same workflow with more autonomy. Codex is the same workflow with OpenAI's distribution.
OpenAI has already moved pricing upward. On April 9, it launched a $100 mid-tier ChatGPT Pro plan with five times the Codex access of the standard Plus tier, explicitly positioned against Anthropic's Claude Max. In the same window, OpenAI paid a reported sum to acquire Astral, the company behind the Python tooling that millions of developers already run. The throughline is visible. OpenAI wants the developer's editor, the developer's runtime, and now the developer's desktop.
The Other Side
Not every developer is convinced that multi-agent desktop control is the future of coding tools.
A section of the Hacker News thread on Thursday's announcement argued that the product is solving a problem most engineers do not have. A single agent that writes careful code with good judgment is more valuable, those commenters wrote, than five agents doing five things at once with no clear ownership of the result. Senior engineers in particular noted that review bandwidth, not generation bandwidth, is the constraint in real software work.
Security researchers flagged a different concern. Handing an agent the ability to click, type, and open arbitrary apps on a Mac collapses the blast radius of a prompt injection into the developer's entire desktop. The LiteLLM supply chain incident in March showed how quickly a compromised tool in a developer's environment can vacuum up cloud credentials. Codex's computer use feature runs with the developer's permissions, which is the point, and also the risk.
Anthropic's supporters pushed back on the framing that Codex has caught up. They pointed out that Claude Code's remote control and coordinator features have been shipping for weeks and that the underlying Opus 4.7 model still leads GPT-5.4 on the benchmarks developers actually cite. A low-grade war, The Bridge Chronicle called it, and noted that at the product level it is not obvious OpenAI is winning on quality. It is winning on speed of feature release.
The Bottom Line
Codex is no longer a code-completion tool that happens to live in a desktop window. As of April 16, it is a workflow orchestrator with computer use, parallel subagents, a browser, memory, image generation, and the beginnings of a plugin economy. That is the same product Anthropic has been quietly building with Claude Code. The difference now is that OpenAI has the distribution to make it the default.
For a working ML engineer or data scientist, the practical change is small and large at the same time. The small part: the next time Codex is asked to refactor a module, it can now open the test runner, read the output, file the JIRA ticket, and update the team's Notion doc without the developer touching any of those windows. The large part: that pattern, once normal, rewrites what a junior engineer's job looks like.
As Altman put it after the 3 million milestone, limits will be reset at every additional million users. Whether the developers who lived through this week's update feel more powerful or more replaceable will decide how long the victory lap lasts.
Sources
- OpenAI takes aim at Anthropic with beefed-up Codex that gives it more power over your desktop — TechCrunch (April 16, 2026)
- OpenAI drastically updates Codex desktop app to use all other apps on your computer, generate images, preview webpages — VentureBeat (April 16, 2026)
- OpenAI Codex Update Adds Computer Use, Image Generation, and Memory on Mac — MacRumors (April 16, 2026)
- OpenAI upgrades Codex with multi-agent workflows and desktop app control to challenge Anthropic's Claude Code — The Tech Portal (April 17, 2026)
- OpenAI ratchets up Codex's agentic capabilities to rival Claude Code — SiliconANGLE (April 16, 2026)
- OpenAI Updates Codex With Computer Use, In-App Browser, Memory, and 90-Plus New Plugins — gHacks Tech News (April 17, 2026)
- OpenAI Codex celebrates 3 million weekly users, CEO Sam Altman resets usage limits — BusinessToday (April 8, 2026)
- Subagents — Codex Developer Documentation (April 16, 2026)
- OpenAI's new $100 ChatGPT Pro plan targets Claude Max with five times the Codex access — The Next Web (March 2026)
- OpenAI turns Codex into super app, Altman takes swipe at Anthropic — The Bridge Chronicle (April 17, 2026)